Why Manual Audits Can’t Keep Up
A quarterly manual accessibility audit checks a sample of pages against WCAG once every three months. Between audits, authors keep publishing new pages, new images, new components and every one of them ships un-audited.
AI doesn’t replace the accessibility expert who interprets edge cases and sets policy. It replaces the impossible task of manually checking every image, every heading, and every form field on every page before every publish catching the repeatable 80% automatically so the expert’s time goes to the 20% that actually needs judgment.
Layer 1 – Pre-Publish Scanning: Catching Issues Before They Ship
An AEM Sidekick extension runs an automated scan against WCAG 2.2 AA criteria the moment an author clicks Preview checking color contrast, heading hierarchy, landmark structure, form labels, and image alt attributes and blocks publish on critical failures rather than flagging them after the fact.
- Contrast ratio checks – computed against the actual rendered CSS, not just an assumption about the design system
- Heading hierarchy validation – flags skipped levels that break screen-reader navigation
- Landmark and ARIA structure – verifies every interactive component has an accessible role and name
Layer 2 – Claude API: Generating Alt Text That’s Actually Descriptive
The single most common WCAG failure is missing or useless alt text. An App Builder action sends each image, along with its surrounding page context, to Claude API to generate alt text that describes what’s actually relevant to a screen-reader user in that context.
Human review stays in the loop: AI-generated alt text is surfaced as a suggestion the author approves or edits never silently written to the DAM without a human seeing it first.
Layer 3 – Automated Remediation: Fixing What’s Safe to Fix Automatically
A subset of WCAG failures are mechanically safe to auto-fix without human review missing lang attributes, empty alt on decorative images, missing form label associations. Everything else contrast, generic alt text, illogical heading order routes to a human review queue.
Where This Runs in the AEM Pipeline
- Author edits a page in Universal Editor and Sidekick extension scans in real time
- Critical failures block publish until resolved or explicitly overridden with a reason logged
- Claude-generated alt text suggestions appear inline for author approval
- A scheduled App Builder job sweeps the full published site weekly, catching drift from component or template updates
- Auto-fixable issues found in the sweep are corrected automatically; everything else creates a remediation ticket
Legal note: Automated scanning and remediation reduces risk but does not by itself constitute a WCAG conformance claim. Regulated industries should still commission periodic third-party manual audits.
Implementation Checklist
- Deploy an AEM Sidekick extension that scans against WCAG 2.2 AA at preview time
- Block publish on critical failures with a logged override path for genuine exceptions
- Build an App Builder action for context-aware alt text generation with mandatory human approval
- Classify WCAG issue types into auto-fixable vs human-review categories before automating anything
- Auto-remediate only the mechanically unambiguous issue types
- Run a scheduled full-site sweep weekly to catch drift from template and component changes
- Route non-auto-fixable issues into a tracked remediation backlog, not a silent log file
- Commission periodic third-party manual audits for regulated content categories regardless of AI coverage
What to Measure
- Pre-publish catch rate – issues blocked at authoring time vs found after publish
- Alt text approval rate – how often authors accept AI-generated text unedited
- Backlog trend – is the human-review queue shrinking or growing month over month
- Time-to-remediate – for issues found in the weekly sweep, how long until fixed
Final Thoughts
The organizations getting ahead of accessibility risk in 2026 aren’t running more audits. They’re catching the repeatable failures before they ever reach a live page, and pointing their actual accessibility experts at the judgment calls AI still can’t make.
Start with alt text and heading structure the two highest-frequency issues before expanding into contrast and ARIA structure checks.
